On 2/13/19 2:33 AM, David Gibson wrote:
> On Tue, Feb 12, 2019 at 08:05:53AM +0100, Cédric Le Goater wrote:
>> On 2/12/19 2:06 AM, David Gibson wrote:
>>> On Mon, Jan 07, 2019 at 07:39:45PM +0100, Cédric Le Goater wrote:
>>>> The IRQ number space of the XIVE and XICS interrupt mode are aligned
>>>> when using the dual interrupt mode for the machine. This means that
>>>> the ICS offset is set to zero in QEMU and that the KVM XICS device
>>>> should be informed of this new value. Unfortunately, there is now way
>>>> to do so and KVM still maintains the XICS_IRQ_BASE (0x1000) offset.
>>>>
>>>> Ignore the lower 4K which are not used under the XICS interrupt
>>>> mode. These IRQ numbers are only claimed by XIVE for the CPU IPIs.
>>>>
>>>> Signed-off-by: Cédric Le Goater <c...@kaod.org>
>>>> ---
>>>> hw/intc/xics_kvm.c | 18 ++++++++++++++++++
>>>> 1 file changed, 18 insertions(+)
>>>>
>>>> diff --git a/hw/intc/xics_kvm.c b/hw/intc/xics_kvm.c
>>>> index 651bbfdf6966..1d21ff217b82 100644
>>>> --- a/hw/intc/xics_kvm.c
>>>> +++ b/hw/intc/xics_kvm.c
>>>> @@ -238,6 +238,15 @@ static void ics_get_kvm_state(ICSState *ics)
>>>> for (i = 0; i < ics->nr_irqs; i++) {
>>>> ICSIRQState *irq = &ics->irqs[i];
>>>>
>>>> + /*
>>>> + * The KVM XICS device considers that the IRQ numbers should
>>>> + * start at XICS_IRQ_BASE (0x1000). Ignore the lower 4K
>>>> + * numbers (only claimed by XIVE for the CPU IPIs).
>>>> + */
>>>> + if (i + ics->offset < XICS_IRQ_BASE) {
>>>> + continue;
>>>> + }
>>>> +
>>>
>>> This seems bogus to me. The guest-visible irq numbers need to line up
>>> between xics and xive mode, yes, but that doesn't mean we need to keep
>>> around a great big array of unused array of ICS irq states, even in
>>> TCG mode.
>>
>> This is because the qirqs[] array is under the machine and shared between
>> both interrupt modes, xics and xive.
>
> I don't see how that follows. ICSIRQState is indexed in terms of the
> ICS source number, not the global irq number, so I don't see why it
> has to match up with the qirq array.
The root cause is the use of spapr->irq->nr_irqs to initialize the ICS
and sPAPRXive object. In case of the 'dual' backend, it covers the full
XIVE IRQ number space (0x2000 today) but XICS only needs 0x1000.
I think we can fix the offset issue by using the appropriate nr_irqs
which should be for the XICS backend : spapr->irq->nr_irqs - ics->offset
I keep in mind the XIVE support for nested guests and I think we will
need to extend the IRQ number space in L1 and have the L2 use a portion
of it (using an offset).
C.
>>>
>>>> kvm_device_access(kernel_xics_fd, KVM_DEV_XICS_GRP_SOURCES,
>>>> i + ics->offset, &state, false, &error_fatal);
>>>>
>>>> @@ -303,6 +312,15 @@ static int ics_set_kvm_state(ICSState *ics, int
>>>> version_id)
>>>> ICSIRQState *irq = &ics->irqs[i];
>>>> int ret;
>>>>
>>>> + /*
>>>> + * The KVM XICS device considers that the IRQ numbers should
>>>> + * start at XICS_IRQ_BASE (0x1000). Ignore the lower 4K
>>>> + * numbers (only claimed by XIVE for the CPU IPIs).
>>>> + */
>>>> + if (i + ics->offset < XICS_IRQ_BASE) {
>>>> + continue;
>>>> + }
>>>> +
>>>> state = irq->server;
>>>> state |= (uint64_t)(irq->saved_priority & KVM_XICS_PRIORITY_MASK)
>>>> << KVM_XICS_PRIORITY_SHIFT;
>>>
>>
>
